Real-Time UAV Sensor Fault Diagnosis Under Mission-Level Navigation Performance Constraints With MR-IGATNet

Abstract

Reliable real-time fault diagnosis of onboard sensors is essential for uncrewed aerial vehicle (UAV) navigation and localization in safety–critical missions. Although recent fault detection and diagnosis (FDD) methods have improved robustness to class imbalance and limited fault data, incipient and noncatastrophic faults under asynchronous multirate sensing and onboard resource constraints remain insufficiently addressed. In this work, we formulate a real-time UAV sensor fault diagnosis problem under mission-level navigation performance constraints, where the operational regime is defined by mission-level navigation error criteria rather than catastrophic mission failure. To address this problem, we propose MR-IGATNet, a multirate multistream diagnosis network that combines InceptionTime window encoding with graph attention (GAT)-based fusion for asynchronous multisensor data. MR-IGATNet adopts a group-wise architecture aligned with feature semantics and sensor update frequencies. It further introduces a dual-level fusion module conditioned on asynchronous update intervals ( $\Delta t$ ), combining graph-attention message passing with query-based pooling (GAT-QP) to integrate information across feature groups and sensor streams. Extensive offline and hardware-in-the-loop (HIL) evaluations show consistent improvements over strong baselines, while real-time deployment results confirm 9.98-ms latency suitable for onboard diagnosis.